AI Investing: A Race Against Expectations
Last Updated on 26 January 2026
The “AI bubble” narrative now dominates headlines. But in markets, a bubble is not just high prices. It is a unique behavioral regime where fear of missing out overrides rational thinking. George Soros described this dynamic within his theory of reflexivity, in his 1987 book The Alchemy of Finance. In this state, rising prices validate the narrative, detaching valuation from reality, and if left untreated- become vulnerable to a shock that forces a violent reset. Price exaggeration itself is not a malfunction. It is the natural mechanism of human perception, which often introduces a gap between perception and reality. Volatility is the way the system digests exaggeration. The critical task is to distinguish between a healthy correction mechanism and an untreated reflexive loop that leads to collapse.
Dot-com comparisons are ubiquitous, yet the nuance lies in the gap between capital invested and the timing of value created. In the late 1990s, the internet was far more promise than product, a dynamic that nonetheless attracted immense capital: between 1996 and 2001, over $2.2 trillion flooded the US market for internet investments, which is roughly $3.9 trillion in 2025 dollars1. This drove expectations to be completely detached from delivery. By 1999, expectations were severely detached, as valuations implied an immediate revolution even though the physical reality was lagging. US e-commerce was a mere 0.64% of retail sales2, and most households were still attached to dial-up connections. Since public-market investors rarely possess a decade-long patience, the failure of prospects to materialize on the implied timeline caused conviction to break, leading to mass liquidation. Ultimately, that spending was not wasted as it financed the fiber backbone and data centers that power today’s web, but the disconnect between the timeline of investment and the timeline of utility created excesses that translated into massive losses for investors.
Today, AI launches from a firmer base. The global infrastructure is built and decades of digital history have aggregated the necessary training data. Capital followed this utility as global investors allocated roughly $2.5 trillion to the sector3. Adoption is broad, with 66% of people in many countries now relying on AI for search and summarization4. Consequently, pricing does not yet seem to be at mania levels: at the 1999 peak, tech leaders traded at 104x trailing earnings5 compared to just about 32x today6 (as measured by the Nasdaq 100). Moreover, in a market where the broader index is already expensive relative to history, the premium paid for AI growth is relatively narrow. However, this dominance creates a structural vulnerability. In 2023, 2024 and 2025, these stocks drove 63%, 53.7% and 42.5% of the total S&P500 index return respectively7,8,9. Although these giants are not pure-play AI startups, their capital expenditures and valuations are now significantly tethered to AI execution. This hints that AI has effectively become a significant part of the market’s Beta. The primary risk is not a speculative bubble in a niche sector but a systemic exposure where the asset class and the broader market are becoming indistinguishable.
Investment success depends on profits scaling fast enough to justify the outlay. As always, the economic logic is inescapable, since total end-user spending must be sufficient to fund the entire capital stack and still yield a decent return. If the addressable market is smaller than the deployed infrastructure, overbuild is a given fact. Beyond this economic limit, the value chain faces a hard physical ceiling: AI workloads demand roughly ten times the power density of traditional cloud computing10, and this capacity cannot scale instantly on demand. The central risk is friction between power and hardware supply and demand. When transformer lead times stretch to years or power siting delays deployment, returns arrive later. In this scenario, capital expenditure outpaces cash generation for years and hints at a classic trap of an over-investment cycle. What keeps investors interested is still the mysticism surrounding those models.
AI delivers real value on real infrastructure. The doubt is not utility, but the timeline of returns against current prevailing expectations, which usually get distorted when retail investors pile in. Have we reached severe FOMO dynamics? Hard to tell. However, markets are pricing a pace of scaling that physical constraints may not be able to support, perhaps implicitly relying on improved chip efficiency in the future. As Soros noticed, public scrutiny can dampen the loop before it turns into a crash, and the result can be a repricing and inter-sector capital flows rather than a violent pop. Either way, a safer posture and solid diversification is to stay anchored to the backend, segments where demand is enforced by necessity and paid for through contracted output; this group currently trades at 16–21x trailing earnings11. That points to grid and regulated power delivery assets, heavy electrical and power hardware, on-site power and energy systems, and thermal, cooling, and physical plant infrastructure, all of which sit far from the front where differentiation is thin and competition compresses returns.
Disclaimer: this article is for informational purposes only and does not constitute financial advice or an endorsement of any specific asset. Investment decisions should be based on one’s own research.
Further reading: Investment Desk at pickel.io.